Parameter Estimation and Implications for Media Planning

We will discuss, in turn, the parameters that describe the attractiveness of each platform, the heterogeneity around their means, the correlations of platform preferences over time

(Σµ) and in intra-day platform usage (Σe). Following that, we present the platform-specific

effects for the tournament characteristics,xkt.

2.5.1. Platform Intercepts and User Heterogeneity

The posterior mean value for pactive is .498, indicating that of our sample, 49.8% of the

registered users are “active,” i.e., they have some predicted probability of accessing soccer content, during the World Cup tournament. Thus, almost half the sample is estimated to be in a spike-at-zero, for each day and each platform, suggesting the necessity for this part of the model. Note, this is consistent with the raw data, in which 61.3% of our sample of 2000 users have no observed usage on the digital platforms.

The general attractiveness to users of the ESPN.com, ESPN3, mobile, and TV platforms

are reflected by the parameter vector µ = (µ1, µ2, µ3, µ4). As we can see in Table 4, the

intercepts for each platform are in the range of -7 to -2 on the logit scale, suggesting baseline (i.e., on non-tournament days) usage of soccer content on the digital platforms is close to zero. This is consistent with the data where, for those users who visit at least once, we see heterogeneous marginal proportions ranging from a low of 2.6% to a high of 100%.

Managerially, it is important to keep in mind that these estimates are based on a sample of users who have been known to use the mobile platform (possibly for non-soccer-related content). Although we have earlier characterized these user as the vanguard of mobile device users, it is interesting that they do not strongly prefer the mobile platform over the others. In fact, the mobile platform has the lowest estimated population mean for the intercept

(µ3 =−6.33), indicating that on days when tournament games are not being played, users

are unlikely to access soccer content on their mobile device.

We also find, unsurprisingly, substantial variation around these population means, with standard errors for the population distribution in the range of 2-2.5, indicating that some users are quite a bit more (or less) likely to access certain platforms,

(diag(Σµ) = (6.358,4.233,5.012)). By contrast, the variances in Σeare quite small (diag(Σe=

(.089, .105, .105, .067)), indicating that after the user’s general propensity to use a platform

is accounted for, there is little residual error in daily usage (other than that driven by the aggregate tournament effects).

2.5.2. Correlation Structure

The correlation structure among the channels across time, and within each day is one of the central areas of interest to media planners. As previously described,we summarize such long-run and daily usage effects in two ways. One is the covariance among users’ propensities

to use each of the platforms over the course of the tournament, which is captured by Σµ,

and the other is the covariance among an individual’s usage of the platforms on a given

day, which is captured by Σe. The estimates of the two covariance matrices are summarized

in Table 4. We observe a strong correlation between ESPN.com and ESPN3 at the long- term level (cross-day posterior mean correlation .795). Thus, heavy users of ESPN.com also tend to be heavy users of ESPN3, which isn’t too surprising given that both platforms are accessed with the same type of device, and so all users who access ESPN.com have the opportunity to access ESPN3. (Note that this correlation estimate is only for active users

Table 4: Estimated Model Parameters: Intercepts and Error Structure

Parameters Mean 2.5% -ile 97.5% -ile

Proportion of Active Users (pactive)

pactive .498 .455 .536

Population Mean for Platform Intercepts (µ)

.com -5.01 -5.01 -4.66

ESPN3 -3.81 -3.82 -3.49

Mobile -6.33 -6.31 -5.96

TV -2.06 -2.06 -1.66

Correlation/Variances for Platform Intercepts (Σµ)

Correlation .com/ESPN3 .795 .749 .835 .com/Mobile .056 -.065 .171 ESPN3/Mobile .028 -.089 .144 Variance .com 6.358 5.279 7.550 ESPN3 4.233 3.326 5.023 Mobile 5.012 3.994 6.340

Correlation/Variances for Daily Error Terms (Σe)

Correlation .com/ESPN3 -.030 -.385 .329 .com/Mobile -.153 -.244 .512 .com/TV .083 -.321 .466 ESPN3/Mobile .114 -.275 .476 ESPN3/TV -.140 -.508 .269 Mobile/TV .016 -.404 .436 Variance .com .089 .051 .149 ESPN3 .105 .062 .172 Mobile .105 .053 .191 TV .067 .037 .117

The posterior mean has been highlighted in bold when the posterior interval does not contain zero.

as non-users are “absorbed” by the spike-at-zero; the estimate of the correlation would be higher were the inactive users included.) In fact, knowing that there are some users who use ESPN.com, but not ESPN3, suggests a relatively easy opportunity for ESPN to expand viewership. Interestingly, we do not find a correlation between ESPN.com and ESPN3 at

the daily level (corr=-.030). Using ESPN.comon a given day does not seem to increase the

chance that the user will watch a streaming video on ESPN3.

Interestingly, the relationship with ESPN’s mobile platform is quite different, and of great business importance given the recent investments made by ESPN (and many other media

long-term or daily correlations with any other platforms. This suggests that mobile usage is not cannibalizing usage of the other platforms.

Finally, we are able to estimate the within-day correlations between TV and the other three platforms, even though we do not directly observe which users are watching TV. Furthermore, the posterior intervals for all the correlations between TV and the other platforms contain zero, suggesting that TV usage on a given day is neither positively or negatively correlated with the use of the digital channels. Interestingly, the posterior mean correlation between ESPN3 and TV is -.140 (2.5%-ile = -.508, 97.5%-ile=2.69), suggesting (directionally) that ESPN3 and TV do compete weakly with each other. This is consistent with the fact that TV and ESPN3 offer very similar content (video of full games). By monitoring this parameter over time, as more data is accumulated, ESPN can keep better track of the relationship between ESPN3 and TV, an issue of key business importance. Summarizing, we find no significant negative correlations between these four channels, sug- gesting that that the content distribution platforms are not at saturation and that new platforms represent an opportunity to generate incremental reach. This is consistent with ESPN’s belief that new platforms do not compete with the old, but allow users to con- sume media at times that they previously could not. Our finding that the mobile platform seems to provide incremental reach, but is still not the most popular platform is consistent with ESPN’s philosophy that users will choose “the best screen available at a given time” (Danaher et al. (2009)).

With our key findings about the relationship between platforms summarized, we now turn to the results that are specific to our case study: the covariates that account for the tournament content.

2.5.3. Tournament effects

As described earlier, the tournament effects include a dummy for whether a given day was on a weekend, the number of games that were played, the number of teams that must “win

or go home” on a given day, and dummies for the US team playing, one of three “culturally significant” teams playing (England, Australia or Mexico) and one of the top three teams playing (Spain, Brazil or Netherlands). The posterior summaries for these coefficients are given in Table 5.

Table 5: Estimated Model Parameters: Tournament Covariates

Parameters Mean 2.5% -ile 97.5% -ile

Weekend .com -.475 -.712 -.250 ESPN3 -.396 -.628 -.163 Mobile .035 -.294 .328 TV .852 .6385 1.082 Number of Games .com 1.370 .842 1.980 ESPN3 -.231 -.842 .472 Mobile .317 -.388 1.012 TV .286 -.190 .705

Number of Teams That Must “Win or Go Home” .com .174 -.304 .707 ESPN3 .095 -.495 .667 Mobile -.145 -.817 .566 TV .081 -.270 .492 US Team Playing .com .338 -.098 .821 ESPN3 .328 -.081 .741 Mobile .699 .213 1.186 TV .523 .227 .935 Canada, Australia or Mexico Playing .com .350 -.054 .726 ESPN3 .028 -.502 .444 Mobile .075 -.425 .583 TV .037 -.309 .415

Top Team Playing

.com .211 -.118 .557

ESPN3 .134 -.314 .563

Mobile .073 -.375 .549

TV .207 -.116 .544

The posterior mean has been highlighted in bold when the posterior interval does not contain zero.

Our results are consistent with the common notion that people are more likely to watch

TV on weekends (β14 = .852) and less likely to go online on weekends (β11 =−.475 and

β12=−.396). (These parameters correspond to users being 2.3 times as likely to watch TV

on the weekend and about .6 times as likely to go online on the weekend.) However, we find

no weekend effect for the mobile platform (β13 = .035). This provides important insight

accessible and used on both weekends and weekdays. While we can only speculate on how mobile will be used in the future, this lack of a day-of-week effect suggests that media plans for the mobile platform will be different than those for TV and online.

Turning to the tournament content itself, we see a (sensible) significant effect for the number of games played on a given day. When there are a large number of games, interested soccer fans increase their usage of ESPN.com substantially as it becomes difficult to follow all the games live on TV, ESPN3 or mobile. Hence, ESPN.com becomes a more attractive platform, while the other platforms are relatively unaffected.

We do not find effects on any of the platforms for the number of teams that must “win or go home.” That is, we do not see any evidence that “clincher” games attract more viewership

on any of the platforms.8

Turning to our set of dummy variables for which teams are playing, we find that all plat- forms, particularly the mobile and TV platforms are more popular when the US team is playing. The estimated parameters indicate users are 2.0 times as likely to access mobile content when the US team is playing and 1.7 times as likely to watch TV.

Interestingly, and perhaps surprising to non-US soccer fans, we find very weak (but positive) effects when a top team (Spain, Brazil or Netherlands) is playing, suggesting that the American audience we observe is more interested in the US team than these top rated soccer teams. For the variable that measures the aforementioned culturally significant teams, we find a weak positive effect for ESPN.com, but not the other platforms. Thus, we find that this US-based audience seems to be most likely to consume content when the US team is playing and is relatively indifferent to which other teams are playing.

Finally, we should note that we are able to achieve good fit with a relatively simple set of covariates describing how many games are being played and who is playing. Note that there are no covariates that describe the “arc” of the tournament; no dummies for the group

8

We thank an anonymous reviewer for the suggestion to include this variable and report the null finding because readers may find it interesting.

stage versus the knockout stage, the final game, etc. While we remain a long way from a complete theory of what makes a game attractive to watch on a particular platform, we note that we are able to capture the aggregate viewership (see Figure 2) with a relatively parsimonious set of covariates.